Fix: Fixed models returning '404 model not found' when using :THINKING suffix

lua/claude_handler.lua

@@ -49,6 +49,10 @@ function _M.handle()
 
     -- Translate Claude request to OpenAI format
     local oai_request = claude_fmt.request_to_openai(claude_body)
+
+    -- Thinking mode handling (operates on OpenAI request format)
+    oai_request, model = e2ee.handle_thinking(oai_request)
+
     local oai_body = cjson.encode(oai_request)
 
     if not is_streaming then

lua/e2ee_discovery.lua

@@ -99,11 +99,18 @@ function _M.resolve_chute_id(model, api_key)
         return model
     end
 
+    -- Strip the :THINKING suffix before map lookup. v1/models returns
+    -- base model IDs only; :THINKING is a proxy-side flag and must not
+    -- leak into the lookup key. Other suffixes (e.g. LoRA names) are
+    -- left intact so they surface as proper errors upstream rather than
+    -- silently resolving to the base chute.
+    local base_model = model:match("^(.-):THINKING$") or model
+
     local now = ngx.now()
 
     -- Check cache
     if model_map and now < model_map_expires then
-        local entry = model_map[model]
+        local entry = model_map[base_model]
         if entry then
             return check_confidential(model, entry)
         end
@@ -114,7 +121,7 @@ function _M.resolve_chute_id(model, api_key)
     if not map then
         -- If we have a stale cache, try it
         if model_map then
-            local entry = model_map[model]
+            local entry = model_map[base_model]
             if entry then
                 return check_confidential(model, entry)
             end
@@ -125,7 +132,7 @@ function _M.resolve_chute_id(model, api_key)
     model_map = map
     model_map_expires = now + MODEL_MAP_TTL
 
-    local entry = map[model]
+    local entry = map[base_model]
     return check_confidential(model, entry)
 end
 

lua/e2ee_handler.lua

@@ -17,6 +17,31 @@ local API_BASE = "https://api.chutes.ai"
 
 local _M = {}
 
+-- Per-model thinking defaults. Mirrors chutes-api's invocation router
+-- (api/invocation/router.py). The /v1/models endpoint exposes a
+-- "reasoning" capability flag in supported_features but not the
+-- per-model default value, so these stay duplicated until the API
+-- grows a dedicated attribute.
+local THINKING_DEFAULT_ON_PREFIXES = {
+    "deepseek-ai/DeepSeek-V3.2-Speciale",
+    "zai-org/GLM-4.7",
+    "moonshotai/Kimi-K2.5",
+}
+local THINKING_DEFAULT_ON_EXACT = {
+    ["deepseek-ai/DeepSeek-V3.2-Speciale"] = true,
+    ["deepseek-ai/DeepSeek-V3.2-Speciale-TEE"] = true,
+    ["zai-org/GLM-4.7"] = true,
+    ["zai-org/GLM-4.7-TEE"] = true,
+    ["moonshotai/Kimi-K2.5"] = true,
+    ["moonshotai/Kimi-K2.5-TEE"] = true,
+}
+local MIMO_PREFIX = "XiaomiMiMo/MiMo-V2-Flash"
+local MIMO_EXACT_TEE = "XiaomiMiMo/MiMo-V2-Flash-TEE"
+
+local function starts_with(s, prefix)
+    return s:sub(1, #prefix) == prefix
+end
+
 --- Extract API key from Authorization header or x-api-key
 function _M.get_api_key()
     local headers = ngx.req.get_headers()
@@ -40,6 +65,68 @@ function _M.get_api_key()
     return key
 end
 
+--- Handle thinking mode for OpenAI format requests
+-- Takes an OpenAI request table, applies thinking logic, and returns the modified table
+-- Also strips :THINKING suffix from model name and returns the model name
+function _M.handle_thinking(oai_request)
+    if not oai_request or not oai_request.model then
+        return oai_request, nil
+    end
+
+    local model = oai_request.model
+    local enable_thinking = false
+    local think_header = ngx.req.get_headers()["X-Enable-Thinking"]
+    if think_header and think_header:lower() == "true" then
+        enable_thinking = true
+    end
+    if model:match(":THINKING$") then
+        model = model:sub(1, -#":THINKING" - 1)
+        oai_request.model = model
+        enable_thinking = true
+    end
+    if enable_thinking then
+        if not oai_request.chat_template_kwargs then
+            oai_request.chat_template_kwargs = {}
+        end
+        oai_request.chat_template_kwargs.thinking = true
+        oai_request.chat_template_kwargs.enable_thinking = true
+    end
+
+    local kwargs = oai_request.chat_template_kwargs
+    if kwargs then
+        -- Normalize the two spellings so downstream sees both.
+        if kwargs.thinking ~= nil and kwargs.enable_thinking == nil then
+            kwargs.enable_thinking = kwargs.thinking
+        end
+        if kwargs.enable_thinking ~= nil and kwargs.thinking == nil then
+            kwargs.thinking = kwargs.enable_thinking
+        end
+
+        if kwargs.thinking == nil then
+            for _, prefix in ipairs(THINKING_DEFAULT_ON_PREFIXES) do
+                if starts_with(model, prefix) then
+                    kwargs.thinking = true
+                    kwargs.enable_thinking = true
+                    break
+                end
+            end
+        end
+
+        if kwargs.thinking == nil and starts_with(model, MIMO_PREFIX) then
+            kwargs.thinking = false
+            kwargs.enable_thinking = false
+        end
+    else
+        if THINKING_DEFAULT_ON_EXACT[model] then
+            oai_request.chat_template_kwargs = { thinking = true, enable_thinking = true }
+        elseif model == MIMO_EXACT_TEE then
+            oai_request.chat_template_kwargs = { thinking = false, enable_thinking = false }
+        end
+    end
+
+    return oai_request, model
+end
+
 --- Send error response
 function _M.send_error(status, message)
     ngx.status = status
@@ -345,6 +432,12 @@ function _M.handle()
         return _M.send_error(400, "missing 'model' field")
     end
 
+    -- Thinking mode handling
+    payload, model = _M.handle_thinking(payload)
+
+    -- Re-encode body with any chat_template_kwargs modifications
+    body = cjson.encode(payload)
+
     local is_streaming = (payload.stream == true)
     local original_path = ngx.var.uri
 

lua/responses_handler.lua

@@ -49,6 +49,10 @@ function _M.handle()
 
     -- Translate Responses request to OpenAI format
     local oai_request = resp_fmt.request_to_openai(resp_body)
+
+    -- Thinking mode handling (operates on OpenAI request format)
+    oai_request, model = e2ee.handle_thinking(oai_request)
+
     local oai_body = cjson.encode(oai_request)
 
     if not is_streaming then